xref: /aosp_15_r20/external/XNNPACK/src/f32-dwconv/gen/up32x25-minmax-avx512f.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f32-dwconv/up-avx512.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2019 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker #include <immintrin.h>
13*4bdc9457SAndroid Build Coastguard Worker 
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/dwconv.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/intrinsics-polyfill.h>
16*4bdc9457SAndroid Build Coastguard Worker 
17*4bdc9457SAndroid Build Coastguard Worker 
xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f(
19*4bdc9457SAndroid Build Coastguard Worker     size_t channels,
20*4bdc9457SAndroid Build Coastguard Worker     size_t output_width,
21*4bdc9457SAndroid Build Coastguard Worker     const float** input,
22*4bdc9457SAndroid Build Coastguard Worker     const float* weights,
23*4bdc9457SAndroid Build Coastguard Worker     float* output,
24*4bdc9457SAndroid Build Coastguard Worker     size_t input_stride,
25*4bdc9457SAndroid Build Coastguard Worker     size_t output_increment,
26*4bdc9457SAndroid Build Coastguard Worker     size_t input_offset,
27*4bdc9457SAndroid Build Coastguard Worker     const float* zero,
28*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29*4bdc9457SAndroid Build Coastguard Worker {
30*4bdc9457SAndroid Build Coastguard Worker   assert(channels != 0);
31*4bdc9457SAndroid Build Coastguard Worker   assert(output_width != 0);
32*4bdc9457SAndroid Build Coastguard Worker 
33*4bdc9457SAndroid Build Coastguard Worker   const __m512 vmax = _mm512_set1_ps(params->scalar.max);
34*4bdc9457SAndroid Build Coastguard Worker   const __m512 vmin = _mm512_set1_ps(params->scalar.min);
35*4bdc9457SAndroid Build Coastguard Worker   do {
36*4bdc9457SAndroid Build Coastguard Worker     const float* i0 = input[0];
37*4bdc9457SAndroid Build Coastguard Worker     assert(i0 != NULL);
38*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i0 != zero) {
39*4bdc9457SAndroid Build Coastguard Worker       i0 = (const float*) ((uintptr_t) i0 + input_offset);
40*4bdc9457SAndroid Build Coastguard Worker     }
41*4bdc9457SAndroid Build Coastguard Worker     const float* i1 = input[1];
42*4bdc9457SAndroid Build Coastguard Worker     assert(i1 != NULL);
43*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i1 != zero) {
44*4bdc9457SAndroid Build Coastguard Worker       i1 = (const float*) ((uintptr_t) i1 + input_offset);
45*4bdc9457SAndroid Build Coastguard Worker     }
46*4bdc9457SAndroid Build Coastguard Worker     const float* i2 = input[2];
47*4bdc9457SAndroid Build Coastguard Worker     assert(i2 != NULL);
48*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i2 != zero) {
49*4bdc9457SAndroid Build Coastguard Worker       i2 = (const float*) ((uintptr_t) i2 + input_offset);
50*4bdc9457SAndroid Build Coastguard Worker     }
51*4bdc9457SAndroid Build Coastguard Worker     const float* i3 = input[3];
52*4bdc9457SAndroid Build Coastguard Worker     assert(i3 != NULL);
53*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i3 != zero) {
54*4bdc9457SAndroid Build Coastguard Worker       i3 = (const float*) ((uintptr_t) i3 + input_offset);
55*4bdc9457SAndroid Build Coastguard Worker     }
56*4bdc9457SAndroid Build Coastguard Worker     const float* i4 = input[4];
57*4bdc9457SAndroid Build Coastguard Worker     assert(i4 != NULL);
58*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i4 != zero) {
59*4bdc9457SAndroid Build Coastguard Worker       i4 = (const float*) ((uintptr_t) i4 + input_offset);
60*4bdc9457SAndroid Build Coastguard Worker     }
61*4bdc9457SAndroid Build Coastguard Worker     const float* i5 = input[5];
62*4bdc9457SAndroid Build Coastguard Worker     assert(i5 != NULL);
63*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i5 != zero) {
64*4bdc9457SAndroid Build Coastguard Worker       i5 = (const float*) ((uintptr_t) i5 + input_offset);
65*4bdc9457SAndroid Build Coastguard Worker     }
66*4bdc9457SAndroid Build Coastguard Worker     const float* i6 = input[6];
67*4bdc9457SAndroid Build Coastguard Worker     assert(i6 != NULL);
68*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i6 != zero) {
69*4bdc9457SAndroid Build Coastguard Worker       i6 = (const float*) ((uintptr_t) i6 + input_offset);
70*4bdc9457SAndroid Build Coastguard Worker     }
71*4bdc9457SAndroid Build Coastguard Worker     const float* i7 = input[7];
72*4bdc9457SAndroid Build Coastguard Worker     assert(i7 != NULL);
73*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i7 != zero) {
74*4bdc9457SAndroid Build Coastguard Worker       i7 = (const float*) ((uintptr_t) i7 + input_offset);
75*4bdc9457SAndroid Build Coastguard Worker     }
76*4bdc9457SAndroid Build Coastguard Worker     const float* i8 = input[8];
77*4bdc9457SAndroid Build Coastguard Worker     assert(i8 != NULL);
78*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i8 != zero) {
79*4bdc9457SAndroid Build Coastguard Worker       i8 = (const float*) ((uintptr_t) i8 + input_offset);
80*4bdc9457SAndroid Build Coastguard Worker     }
81*4bdc9457SAndroid Build Coastguard Worker     const float* i9 = input[9];
82*4bdc9457SAndroid Build Coastguard Worker     assert(i9 != NULL);
83*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i9 != zero) {
84*4bdc9457SAndroid Build Coastguard Worker       i9 = (const float*) ((uintptr_t) i9 + input_offset);
85*4bdc9457SAndroid Build Coastguard Worker     }
86*4bdc9457SAndroid Build Coastguard Worker     const float* i10 = input[10];
87*4bdc9457SAndroid Build Coastguard Worker     assert(i10 != NULL);
88*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i10 != zero) {
89*4bdc9457SAndroid Build Coastguard Worker       i10 = (const float*) ((uintptr_t) i10 + input_offset);
90*4bdc9457SAndroid Build Coastguard Worker     }
91*4bdc9457SAndroid Build Coastguard Worker     const float* i11 = input[11];
92*4bdc9457SAndroid Build Coastguard Worker     assert(i11 != NULL);
93*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i11 != zero) {
94*4bdc9457SAndroid Build Coastguard Worker       i11 = (const float*) ((uintptr_t) i11 + input_offset);
95*4bdc9457SAndroid Build Coastguard Worker     }
96*4bdc9457SAndroid Build Coastguard Worker     const float* i12 = input[12];
97*4bdc9457SAndroid Build Coastguard Worker     assert(i12 != NULL);
98*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i12 != zero) {
99*4bdc9457SAndroid Build Coastguard Worker       i12 = (const float*) ((uintptr_t) i12 + input_offset);
100*4bdc9457SAndroid Build Coastguard Worker     }
101*4bdc9457SAndroid Build Coastguard Worker     const float* i13 = input[13];
102*4bdc9457SAndroid Build Coastguard Worker     assert(i13 != NULL);
103*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i13 != zero) {
104*4bdc9457SAndroid Build Coastguard Worker       i13 = (const float*) ((uintptr_t) i13 + input_offset);
105*4bdc9457SAndroid Build Coastguard Worker     }
106*4bdc9457SAndroid Build Coastguard Worker     const float* i14 = input[14];
107*4bdc9457SAndroid Build Coastguard Worker     assert(i14 != NULL);
108*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i14 != zero) {
109*4bdc9457SAndroid Build Coastguard Worker       i14 = (const float*) ((uintptr_t) i14 + input_offset);
110*4bdc9457SAndroid Build Coastguard Worker     }
111*4bdc9457SAndroid Build Coastguard Worker     const float* i15 = input[15];
112*4bdc9457SAndroid Build Coastguard Worker     assert(i15 != NULL);
113*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i15 != zero) {
114*4bdc9457SAndroid Build Coastguard Worker       i15 = (const float*) ((uintptr_t) i15 + input_offset);
115*4bdc9457SAndroid Build Coastguard Worker     }
116*4bdc9457SAndroid Build Coastguard Worker     const float* i16 = input[16];
117*4bdc9457SAndroid Build Coastguard Worker     assert(i16 != NULL);
118*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i16 != zero) {
119*4bdc9457SAndroid Build Coastguard Worker       i16 = (const float*) ((uintptr_t) i16 + input_offset);
120*4bdc9457SAndroid Build Coastguard Worker     }
121*4bdc9457SAndroid Build Coastguard Worker     const float* i17 = input[17];
122*4bdc9457SAndroid Build Coastguard Worker     assert(i17 != NULL);
123*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i17 != zero) {
124*4bdc9457SAndroid Build Coastguard Worker       i17 = (const float*) ((uintptr_t) i17 + input_offset);
125*4bdc9457SAndroid Build Coastguard Worker     }
126*4bdc9457SAndroid Build Coastguard Worker     const float* i18 = input[18];
127*4bdc9457SAndroid Build Coastguard Worker     assert(i18 != NULL);
128*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i18 != zero) {
129*4bdc9457SAndroid Build Coastguard Worker       i18 = (const float*) ((uintptr_t) i18 + input_offset);
130*4bdc9457SAndroid Build Coastguard Worker     }
131*4bdc9457SAndroid Build Coastguard Worker     const float* i19 = input[19];
132*4bdc9457SAndroid Build Coastguard Worker     assert(i19 != NULL);
133*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i19 != zero) {
134*4bdc9457SAndroid Build Coastguard Worker       i19 = (const float*) ((uintptr_t) i19 + input_offset);
135*4bdc9457SAndroid Build Coastguard Worker     }
136*4bdc9457SAndroid Build Coastguard Worker     const float* i20 = input[20];
137*4bdc9457SAndroid Build Coastguard Worker     assert(i20 != NULL);
138*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i20 != zero) {
139*4bdc9457SAndroid Build Coastguard Worker       i20 = (const float*) ((uintptr_t) i20 + input_offset);
140*4bdc9457SAndroid Build Coastguard Worker     }
141*4bdc9457SAndroid Build Coastguard Worker     const float* i21 = input[21];
142*4bdc9457SAndroid Build Coastguard Worker     assert(i21 != NULL);
143*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i21 != zero) {
144*4bdc9457SAndroid Build Coastguard Worker       i21 = (const float*) ((uintptr_t) i21 + input_offset);
145*4bdc9457SAndroid Build Coastguard Worker     }
146*4bdc9457SAndroid Build Coastguard Worker     const float* i22 = input[22];
147*4bdc9457SAndroid Build Coastguard Worker     assert(i22 != NULL);
148*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i22 != zero) {
149*4bdc9457SAndroid Build Coastguard Worker       i22 = (const float*) ((uintptr_t) i22 + input_offset);
150*4bdc9457SAndroid Build Coastguard Worker     }
151*4bdc9457SAndroid Build Coastguard Worker     const float* i23 = input[23];
152*4bdc9457SAndroid Build Coastguard Worker     assert(i23 != NULL);
153*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i23 != zero) {
154*4bdc9457SAndroid Build Coastguard Worker       i23 = (const float*) ((uintptr_t) i23 + input_offset);
155*4bdc9457SAndroid Build Coastguard Worker     }
156*4bdc9457SAndroid Build Coastguard Worker     const float* i24 = input[24];
157*4bdc9457SAndroid Build Coastguard Worker     assert(i24 != NULL);
158*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(i24 != zero) {
159*4bdc9457SAndroid Build Coastguard Worker       i24 = (const float*) ((uintptr_t) i24 + input_offset);
160*4bdc9457SAndroid Build Coastguard Worker     }
161*4bdc9457SAndroid Build Coastguard Worker     input = (const float**) ((uintptr_t) input + input_stride);
162*4bdc9457SAndroid Build Coastguard Worker 
163*4bdc9457SAndroid Build Coastguard Worker     size_t c = channels;
164*4bdc9457SAndroid Build Coastguard Worker     const float* w = weights;
165*4bdc9457SAndroid Build Coastguard Worker     for (; c >= 32; c -= 32) {
166*4bdc9457SAndroid Build Coastguard Worker       __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
167*4bdc9457SAndroid Build Coastguard Worker       __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16);
168*4bdc9457SAndroid Build Coastguard Worker 
169*4bdc9457SAndroid Build Coastguard Worker 
170*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
171*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16);
172*4bdc9457SAndroid Build Coastguard Worker       i0 += 32;
173*4bdc9457SAndroid Build Coastguard Worker 
174*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32);
175*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48);
176*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
177*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
178*4bdc9457SAndroid Build Coastguard Worker 
179*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
180*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16);
181*4bdc9457SAndroid Build Coastguard Worker       i1 += 32;
182*4bdc9457SAndroid Build Coastguard Worker 
183*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64);
184*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80);
185*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
186*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
187*4bdc9457SAndroid Build Coastguard Worker 
188*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
189*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16);
190*4bdc9457SAndroid Build Coastguard Worker       i2 += 32;
191*4bdc9457SAndroid Build Coastguard Worker 
192*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96);
193*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112);
194*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
195*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
196*4bdc9457SAndroid Build Coastguard Worker 
197*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3);
198*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi3xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i3 + 16);
199*4bdc9457SAndroid Build Coastguard Worker       i3 += 32;
200*4bdc9457SAndroid Build Coastguard Worker 
201*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128);
202*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk3xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 144);
203*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
204*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
205*4bdc9457SAndroid Build Coastguard Worker 
206*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4);
207*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi4xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i4 + 16);
208*4bdc9457SAndroid Build Coastguard Worker       i4 += 32;
209*4bdc9457SAndroid Build Coastguard Worker 
210*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 160);
211*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk4xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 176);
212*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc0123456789ABCDEFp0);
213*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
214*4bdc9457SAndroid Build Coastguard Worker 
215*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5);
216*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi5xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i5 + 16);
217*4bdc9457SAndroid Build Coastguard Worker       i5 += 32;
218*4bdc9457SAndroid Build Coastguard Worker 
219*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192);
220*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk5xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 208);
221*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc0123456789ABCDEFp0);
222*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
223*4bdc9457SAndroid Build Coastguard Worker 
224*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6);
225*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi6xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i6 + 16);
226*4bdc9457SAndroid Build Coastguard Worker       i6 += 32;
227*4bdc9457SAndroid Build Coastguard Worker 
228*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 224);
229*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk6xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 240);
230*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc0123456789ABCDEFp0);
231*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
232*4bdc9457SAndroid Build Coastguard Worker 
233*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7);
234*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi7xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i7 + 16);
235*4bdc9457SAndroid Build Coastguard Worker       i7 += 32;
236*4bdc9457SAndroid Build Coastguard Worker 
237*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 256);
238*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk7xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 272);
239*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF, vacc0123456789ABCDEFp0);
240*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
241*4bdc9457SAndroid Build Coastguard Worker 
242*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8);
243*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi8xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i8 + 16);
244*4bdc9457SAndroid Build Coastguard Worker       i8 += 32;
245*4bdc9457SAndroid Build Coastguard Worker 
246*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288);
247*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk8xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 304);
248*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc0123456789ABCDEFp0);
249*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
250*4bdc9457SAndroid Build Coastguard Worker 
251*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi9x0123456789ABCDEF = _mm512_loadu_ps(i9);
252*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi9xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i9 + 16);
253*4bdc9457SAndroid Build Coastguard Worker       i9 += 32;
254*4bdc9457SAndroid Build Coastguard Worker 
255*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk9x0123456789ABCDEF = _mm512_load_ps(w + 320);
256*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk9xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 336);
257*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF, vacc0123456789ABCDEFp0);
258*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi9xGHIJKLMNOPQRSTUV, vk9xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
259*4bdc9457SAndroid Build Coastguard Worker 
260*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10);
261*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi10xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i10 + 16);
262*4bdc9457SAndroid Build Coastguard Worker       i10 += 32;
263*4bdc9457SAndroid Build Coastguard Worker 
264*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk10x0123456789ABCDEF = _mm512_load_ps(w + 352);
265*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk10xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 368);
266*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF, vacc0123456789ABCDEFp0);
267*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi10xGHIJKLMNOPQRSTUV, vk10xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
268*4bdc9457SAndroid Build Coastguard Worker 
269*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi11x0123456789ABCDEF = _mm512_loadu_ps(i11);
270*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi11xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i11 + 16);
271*4bdc9457SAndroid Build Coastguard Worker       i11 += 32;
272*4bdc9457SAndroid Build Coastguard Worker 
273*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 384);
274*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk11xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 400);
275*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF, vacc0123456789ABCDEFp0);
276*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi11xGHIJKLMNOPQRSTUV, vk11xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
277*4bdc9457SAndroid Build Coastguard Worker 
278*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi12x0123456789ABCDEF = _mm512_loadu_ps(i12);
279*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi12xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i12 + 16);
280*4bdc9457SAndroid Build Coastguard Worker       i12 += 32;
281*4bdc9457SAndroid Build Coastguard Worker 
282*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk12x0123456789ABCDEF = _mm512_load_ps(w + 416);
283*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk12xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 432);
284*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF, vacc0123456789ABCDEFp0);
285*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi12xGHIJKLMNOPQRSTUV, vk12xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
286*4bdc9457SAndroid Build Coastguard Worker 
287*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi13x0123456789ABCDEF = _mm512_loadu_ps(i13);
288*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi13xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i13 + 16);
289*4bdc9457SAndroid Build Coastguard Worker       i13 += 32;
290*4bdc9457SAndroid Build Coastguard Worker 
291*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk13x0123456789ABCDEF = _mm512_load_ps(w + 448);
292*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk13xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 464);
293*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF, vacc0123456789ABCDEFp0);
294*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi13xGHIJKLMNOPQRSTUV, vk13xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
295*4bdc9457SAndroid Build Coastguard Worker 
296*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi14x0123456789ABCDEF = _mm512_loadu_ps(i14);
297*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi14xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i14 + 16);
298*4bdc9457SAndroid Build Coastguard Worker       i14 += 32;
299*4bdc9457SAndroid Build Coastguard Worker 
300*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk14x0123456789ABCDEF = _mm512_load_ps(w + 480);
301*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk14xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 496);
302*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF, vacc0123456789ABCDEFp0);
303*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi14xGHIJKLMNOPQRSTUV, vk14xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
304*4bdc9457SAndroid Build Coastguard Worker 
305*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi15x0123456789ABCDEF = _mm512_loadu_ps(i15);
306*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi15xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i15 + 16);
307*4bdc9457SAndroid Build Coastguard Worker       i15 += 32;
308*4bdc9457SAndroid Build Coastguard Worker 
309*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk15x0123456789ABCDEF = _mm512_load_ps(w + 512);
310*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk15xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 528);
311*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF, vacc0123456789ABCDEFp0);
312*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi15xGHIJKLMNOPQRSTUV, vk15xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
313*4bdc9457SAndroid Build Coastguard Worker 
314*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi16x0123456789ABCDEF = _mm512_loadu_ps(i16);
315*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi16xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i16 + 16);
316*4bdc9457SAndroid Build Coastguard Worker       i16 += 32;
317*4bdc9457SAndroid Build Coastguard Worker 
318*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk16x0123456789ABCDEF = _mm512_load_ps(w + 544);
319*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk16xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 560);
320*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF, vacc0123456789ABCDEFp0);
321*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi16xGHIJKLMNOPQRSTUV, vk16xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
322*4bdc9457SAndroid Build Coastguard Worker 
323*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi17x0123456789ABCDEF = _mm512_loadu_ps(i17);
324*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi17xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i17 + 16);
325*4bdc9457SAndroid Build Coastguard Worker       i17 += 32;
326*4bdc9457SAndroid Build Coastguard Worker 
327*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk17x0123456789ABCDEF = _mm512_load_ps(w + 576);
328*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk17xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 592);
329*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF, vacc0123456789ABCDEFp0);
330*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi17xGHIJKLMNOPQRSTUV, vk17xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
331*4bdc9457SAndroid Build Coastguard Worker 
332*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi18x0123456789ABCDEF = _mm512_loadu_ps(i18);
333*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi18xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i18 + 16);
334*4bdc9457SAndroid Build Coastguard Worker       i18 += 32;
335*4bdc9457SAndroid Build Coastguard Worker 
336*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk18x0123456789ABCDEF = _mm512_load_ps(w + 608);
337*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk18xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 624);
338*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF, vacc0123456789ABCDEFp0);
339*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi18xGHIJKLMNOPQRSTUV, vk18xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
340*4bdc9457SAndroid Build Coastguard Worker 
341*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi19x0123456789ABCDEF = _mm512_loadu_ps(i19);
342*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi19xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i19 + 16);
343*4bdc9457SAndroid Build Coastguard Worker       i19 += 32;
344*4bdc9457SAndroid Build Coastguard Worker 
345*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk19x0123456789ABCDEF = _mm512_load_ps(w + 640);
346*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk19xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 656);
347*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF, vacc0123456789ABCDEFp0);
348*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi19xGHIJKLMNOPQRSTUV, vk19xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
349*4bdc9457SAndroid Build Coastguard Worker 
350*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi20x0123456789ABCDEF = _mm512_loadu_ps(i20);
351*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi20xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i20 + 16);
352*4bdc9457SAndroid Build Coastguard Worker       i20 += 32;
353*4bdc9457SAndroid Build Coastguard Worker 
354*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk20x0123456789ABCDEF = _mm512_load_ps(w + 672);
355*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk20xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 688);
356*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF, vacc0123456789ABCDEFp0);
357*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi20xGHIJKLMNOPQRSTUV, vk20xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
358*4bdc9457SAndroid Build Coastguard Worker 
359*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi21x0123456789ABCDEF = _mm512_loadu_ps(i21);
360*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi21xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i21 + 16);
361*4bdc9457SAndroid Build Coastguard Worker       i21 += 32;
362*4bdc9457SAndroid Build Coastguard Worker 
363*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk21x0123456789ABCDEF = _mm512_load_ps(w + 704);
364*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk21xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 720);
365*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF, vacc0123456789ABCDEFp0);
366*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi21xGHIJKLMNOPQRSTUV, vk21xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
367*4bdc9457SAndroid Build Coastguard Worker 
368*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi22x0123456789ABCDEF = _mm512_loadu_ps(i22);
369*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi22xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i22 + 16);
370*4bdc9457SAndroid Build Coastguard Worker       i22 += 32;
371*4bdc9457SAndroid Build Coastguard Worker 
372*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk22x0123456789ABCDEF = _mm512_load_ps(w + 736);
373*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk22xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 752);
374*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF, vacc0123456789ABCDEFp0);
375*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi22xGHIJKLMNOPQRSTUV, vk22xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
376*4bdc9457SAndroid Build Coastguard Worker 
377*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi23x0123456789ABCDEF = _mm512_loadu_ps(i23);
378*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi23xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i23 + 16);
379*4bdc9457SAndroid Build Coastguard Worker       i23 += 32;
380*4bdc9457SAndroid Build Coastguard Worker 
381*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk23x0123456789ABCDEF = _mm512_load_ps(w + 768);
382*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk23xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 784);
383*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF, vacc0123456789ABCDEFp0);
384*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi23xGHIJKLMNOPQRSTUV, vk23xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
385*4bdc9457SAndroid Build Coastguard Worker 
386*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi24x0123456789ABCDEF = _mm512_loadu_ps(i24);
387*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi24xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i24 + 16);
388*4bdc9457SAndroid Build Coastguard Worker       i24 += 32;
389*4bdc9457SAndroid Build Coastguard Worker 
390*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk24x0123456789ABCDEF = _mm512_load_ps(w + 800);
391*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk24xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 816);
392*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF, vacc0123456789ABCDEFp0);
393*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi24xGHIJKLMNOPQRSTUV, vk24xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
394*4bdc9457SAndroid Build Coastguard Worker 
395*4bdc9457SAndroid Build Coastguard Worker       w += 832;
396*4bdc9457SAndroid Build Coastguard Worker 
397*4bdc9457SAndroid Build Coastguard Worker 
398*4bdc9457SAndroid Build Coastguard Worker       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
399*4bdc9457SAndroid Build Coastguard Worker       __m512 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUVp0, vmin);
400*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
401*4bdc9457SAndroid Build Coastguard Worker       vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vmax);
402*4bdc9457SAndroid Build Coastguard Worker 
403*4bdc9457SAndroid Build Coastguard Worker       _mm512_storeu_ps(output, vacc0123456789ABCDEF);
404*4bdc9457SAndroid Build Coastguard Worker       _mm512_storeu_ps(output + 16, vaccGHIJKLMNOPQRSTUV);
405*4bdc9457SAndroid Build Coastguard Worker       output += 32;
406*4bdc9457SAndroid Build Coastguard Worker     }
407*4bdc9457SAndroid Build Coastguard Worker     for (; c >= 16; c -= 16) {
408*4bdc9457SAndroid Build Coastguard Worker       __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
409*4bdc9457SAndroid Build Coastguard Worker 
410*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
411*4bdc9457SAndroid Build Coastguard Worker       i0 += 16;
412*4bdc9457SAndroid Build Coastguard Worker 
413*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32);
414*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
415*4bdc9457SAndroid Build Coastguard Worker 
416*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
417*4bdc9457SAndroid Build Coastguard Worker       i1 += 16;
418*4bdc9457SAndroid Build Coastguard Worker 
419*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64);
420*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
421*4bdc9457SAndroid Build Coastguard Worker 
422*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
423*4bdc9457SAndroid Build Coastguard Worker       i2 += 16;
424*4bdc9457SAndroid Build Coastguard Worker 
425*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96);
426*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
427*4bdc9457SAndroid Build Coastguard Worker 
428*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3);
429*4bdc9457SAndroid Build Coastguard Worker       i3 += 16;
430*4bdc9457SAndroid Build Coastguard Worker 
431*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128);
432*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
433*4bdc9457SAndroid Build Coastguard Worker 
434*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4);
435*4bdc9457SAndroid Build Coastguard Worker       i4 += 16;
436*4bdc9457SAndroid Build Coastguard Worker 
437*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 160);
438*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc0123456789ABCDEFp0);
439*4bdc9457SAndroid Build Coastguard Worker 
440*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5);
441*4bdc9457SAndroid Build Coastguard Worker       i5 += 16;
442*4bdc9457SAndroid Build Coastguard Worker 
443*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192);
444*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc0123456789ABCDEFp0);
445*4bdc9457SAndroid Build Coastguard Worker 
446*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6);
447*4bdc9457SAndroid Build Coastguard Worker       i6 += 16;
448*4bdc9457SAndroid Build Coastguard Worker 
449*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 224);
450*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc0123456789ABCDEFp0);
451*4bdc9457SAndroid Build Coastguard Worker 
452*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7);
453*4bdc9457SAndroid Build Coastguard Worker       i7 += 16;
454*4bdc9457SAndroid Build Coastguard Worker 
455*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 256);
456*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF, vacc0123456789ABCDEFp0);
457*4bdc9457SAndroid Build Coastguard Worker 
458*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8);
459*4bdc9457SAndroid Build Coastguard Worker       i8 += 16;
460*4bdc9457SAndroid Build Coastguard Worker 
461*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288);
462*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc0123456789ABCDEFp0);
463*4bdc9457SAndroid Build Coastguard Worker 
464*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi9x0123456789ABCDEF = _mm512_loadu_ps(i9);
465*4bdc9457SAndroid Build Coastguard Worker       i9 += 16;
466*4bdc9457SAndroid Build Coastguard Worker 
467*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk9x0123456789ABCDEF = _mm512_load_ps(w + 320);
468*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF, vacc0123456789ABCDEFp0);
469*4bdc9457SAndroid Build Coastguard Worker 
470*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10);
471*4bdc9457SAndroid Build Coastguard Worker       i10 += 16;
472*4bdc9457SAndroid Build Coastguard Worker 
473*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk10x0123456789ABCDEF = _mm512_load_ps(w + 352);
474*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF, vacc0123456789ABCDEFp0);
475*4bdc9457SAndroid Build Coastguard Worker 
476*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi11x0123456789ABCDEF = _mm512_loadu_ps(i11);
477*4bdc9457SAndroid Build Coastguard Worker       i11 += 16;
478*4bdc9457SAndroid Build Coastguard Worker 
479*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 384);
480*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF, vacc0123456789ABCDEFp0);
481*4bdc9457SAndroid Build Coastguard Worker 
482*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi12x0123456789ABCDEF = _mm512_loadu_ps(i12);
483*4bdc9457SAndroid Build Coastguard Worker       i12 += 16;
484*4bdc9457SAndroid Build Coastguard Worker 
485*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk12x0123456789ABCDEF = _mm512_load_ps(w + 416);
486*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF, vacc0123456789ABCDEFp0);
487*4bdc9457SAndroid Build Coastguard Worker 
488*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi13x0123456789ABCDEF = _mm512_loadu_ps(i13);
489*4bdc9457SAndroid Build Coastguard Worker       i13 += 16;
490*4bdc9457SAndroid Build Coastguard Worker 
491*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk13x0123456789ABCDEF = _mm512_load_ps(w + 448);
492*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF, vacc0123456789ABCDEFp0);
493*4bdc9457SAndroid Build Coastguard Worker 
494*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi14x0123456789ABCDEF = _mm512_loadu_ps(i14);
495*4bdc9457SAndroid Build Coastguard Worker       i14 += 16;
496*4bdc9457SAndroid Build Coastguard Worker 
497*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk14x0123456789ABCDEF = _mm512_load_ps(w + 480);
498*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF, vacc0123456789ABCDEFp0);
499*4bdc9457SAndroid Build Coastguard Worker 
500*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi15x0123456789ABCDEF = _mm512_loadu_ps(i15);
501*4bdc9457SAndroid Build Coastguard Worker       i15 += 16;
502*4bdc9457SAndroid Build Coastguard Worker 
503*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk15x0123456789ABCDEF = _mm512_load_ps(w + 512);
504*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF, vacc0123456789ABCDEFp0);
505*4bdc9457SAndroid Build Coastguard Worker 
506*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi16x0123456789ABCDEF = _mm512_loadu_ps(i16);
507*4bdc9457SAndroid Build Coastguard Worker       i16 += 16;
508*4bdc9457SAndroid Build Coastguard Worker 
509*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk16x0123456789ABCDEF = _mm512_load_ps(w + 544);
510*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF, vacc0123456789ABCDEFp0);
511*4bdc9457SAndroid Build Coastguard Worker 
512*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi17x0123456789ABCDEF = _mm512_loadu_ps(i17);
513*4bdc9457SAndroid Build Coastguard Worker       i17 += 16;
514*4bdc9457SAndroid Build Coastguard Worker 
515*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk17x0123456789ABCDEF = _mm512_load_ps(w + 576);
516*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF, vacc0123456789ABCDEFp0);
517*4bdc9457SAndroid Build Coastguard Worker 
518*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi18x0123456789ABCDEF = _mm512_loadu_ps(i18);
519*4bdc9457SAndroid Build Coastguard Worker       i18 += 16;
520*4bdc9457SAndroid Build Coastguard Worker 
521*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk18x0123456789ABCDEF = _mm512_load_ps(w + 608);
522*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF, vacc0123456789ABCDEFp0);
523*4bdc9457SAndroid Build Coastguard Worker 
524*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi19x0123456789ABCDEF = _mm512_loadu_ps(i19);
525*4bdc9457SAndroid Build Coastguard Worker       i19 += 16;
526*4bdc9457SAndroid Build Coastguard Worker 
527*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk19x0123456789ABCDEF = _mm512_load_ps(w + 640);
528*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF, vacc0123456789ABCDEFp0);
529*4bdc9457SAndroid Build Coastguard Worker 
530*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi20x0123456789ABCDEF = _mm512_loadu_ps(i20);
531*4bdc9457SAndroid Build Coastguard Worker       i20 += 16;
532*4bdc9457SAndroid Build Coastguard Worker 
533*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk20x0123456789ABCDEF = _mm512_load_ps(w + 672);
534*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF, vacc0123456789ABCDEFp0);
535*4bdc9457SAndroid Build Coastguard Worker 
536*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi21x0123456789ABCDEF = _mm512_loadu_ps(i21);
537*4bdc9457SAndroid Build Coastguard Worker       i21 += 16;
538*4bdc9457SAndroid Build Coastguard Worker 
539*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk21x0123456789ABCDEF = _mm512_load_ps(w + 704);
540*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF, vacc0123456789ABCDEFp0);
541*4bdc9457SAndroid Build Coastguard Worker 
542*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi22x0123456789ABCDEF = _mm512_loadu_ps(i22);
543*4bdc9457SAndroid Build Coastguard Worker       i22 += 16;
544*4bdc9457SAndroid Build Coastguard Worker 
545*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk22x0123456789ABCDEF = _mm512_load_ps(w + 736);
546*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF, vacc0123456789ABCDEFp0);
547*4bdc9457SAndroid Build Coastguard Worker 
548*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi23x0123456789ABCDEF = _mm512_loadu_ps(i23);
549*4bdc9457SAndroid Build Coastguard Worker       i23 += 16;
550*4bdc9457SAndroid Build Coastguard Worker 
551*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk23x0123456789ABCDEF = _mm512_load_ps(w + 768);
552*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF, vacc0123456789ABCDEFp0);
553*4bdc9457SAndroid Build Coastguard Worker 
554*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi24x0123456789ABCDEF = _mm512_loadu_ps(i24);
555*4bdc9457SAndroid Build Coastguard Worker       i24 += 16;
556*4bdc9457SAndroid Build Coastguard Worker 
557*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk24x0123456789ABCDEF = _mm512_load_ps(w + 800);
558*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF, vacc0123456789ABCDEFp0);
559*4bdc9457SAndroid Build Coastguard Worker 
560*4bdc9457SAndroid Build Coastguard Worker       w += 16;
561*4bdc9457SAndroid Build Coastguard Worker 
562*4bdc9457SAndroid Build Coastguard Worker 
563*4bdc9457SAndroid Build Coastguard Worker       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
564*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
565*4bdc9457SAndroid Build Coastguard Worker 
566*4bdc9457SAndroid Build Coastguard Worker       _mm512_storeu_ps(output, vacc0123456789ABCDEF);
567*4bdc9457SAndroid Build Coastguard Worker       output += 16;
568*4bdc9457SAndroid Build Coastguard Worker     }
569*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNLIKELY(c != 0) {
570*4bdc9457SAndroid Build Coastguard Worker       assert(c >= 1);
571*4bdc9457SAndroid Build Coastguard Worker       assert(c <= 16);
572*4bdc9457SAndroid Build Coastguard Worker       // Prepare mask for valid 32-bit elements (depends on nc).
573*4bdc9457SAndroid Build Coastguard Worker       const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << c) - UINT32_C(1)));
574*4bdc9457SAndroid Build Coastguard Worker 
575*4bdc9457SAndroid Build Coastguard Worker       __m512 vacc0123456789ABCDEFp0 = _mm512_maskz_loadu_ps(vmask, w);
576*4bdc9457SAndroid Build Coastguard Worker 
577*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0);
578*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 32);
579*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
580*4bdc9457SAndroid Build Coastguard Worker 
581*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i1);
582*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 64);
583*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
584*4bdc9457SAndroid Build Coastguard Worker 
585*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i2);
586*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 96);
587*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
588*4bdc9457SAndroid Build Coastguard Worker 
589*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3);
590*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 128);
591*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
592*4bdc9457SAndroid Build Coastguard Worker 
593*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4);
594*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 160);
595*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc0123456789ABCDEFp0);
596*4bdc9457SAndroid Build Coastguard Worker 
597*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i5);
598*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 192);
599*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc0123456789ABCDEFp0);
600*4bdc9457SAndroid Build Coastguard Worker 
601*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6);
602*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 224);
603*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc0123456789ABCDEFp0);
604*4bdc9457SAndroid Build Coastguard Worker 
605*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi7x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i7);
606*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk7x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 256);
607*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF, vacc0123456789ABCDEFp0);
608*4bdc9457SAndroid Build Coastguard Worker 
609*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8);
610*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 288);
611*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc0123456789ABCDEFp0);
612*4bdc9457SAndroid Build Coastguard Worker 
613*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi9x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i9);
614*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk9x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 320);
615*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF, vacc0123456789ABCDEFp0);
616*4bdc9457SAndroid Build Coastguard Worker 
617*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi10x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i10);
618*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk10x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 352);
619*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF, vacc0123456789ABCDEFp0);
620*4bdc9457SAndroid Build Coastguard Worker 
621*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi11x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i11);
622*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk11x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 384);
623*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF, vacc0123456789ABCDEFp0);
624*4bdc9457SAndroid Build Coastguard Worker 
625*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi12x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i12);
626*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk12x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 416);
627*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF, vacc0123456789ABCDEFp0);
628*4bdc9457SAndroid Build Coastguard Worker 
629*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi13x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i13);
630*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk13x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 448);
631*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF, vacc0123456789ABCDEFp0);
632*4bdc9457SAndroid Build Coastguard Worker 
633*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi14x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i14);
634*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk14x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 480);
635*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF, vacc0123456789ABCDEFp0);
636*4bdc9457SAndroid Build Coastguard Worker 
637*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi15x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i15);
638*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk15x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 512);
639*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF, vacc0123456789ABCDEFp0);
640*4bdc9457SAndroid Build Coastguard Worker 
641*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi16x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i16);
642*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk16x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 544);
643*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF, vacc0123456789ABCDEFp0);
644*4bdc9457SAndroid Build Coastguard Worker 
645*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi17x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i17);
646*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk17x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 576);
647*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF, vacc0123456789ABCDEFp0);
648*4bdc9457SAndroid Build Coastguard Worker 
649*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi18x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i18);
650*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk18x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 608);
651*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF, vacc0123456789ABCDEFp0);
652*4bdc9457SAndroid Build Coastguard Worker 
653*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi19x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i19);
654*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk19x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 640);
655*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF, vacc0123456789ABCDEFp0);
656*4bdc9457SAndroid Build Coastguard Worker 
657*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi20x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i20);
658*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk20x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 672);
659*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF, vacc0123456789ABCDEFp0);
660*4bdc9457SAndroid Build Coastguard Worker 
661*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi21x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i21);
662*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk21x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 704);
663*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF, vacc0123456789ABCDEFp0);
664*4bdc9457SAndroid Build Coastguard Worker 
665*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi22x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i22);
666*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk22x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 736);
667*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF, vacc0123456789ABCDEFp0);
668*4bdc9457SAndroid Build Coastguard Worker 
669*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi23x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i23);
670*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk23x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 768);
671*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF, vacc0123456789ABCDEFp0);
672*4bdc9457SAndroid Build Coastguard Worker 
673*4bdc9457SAndroid Build Coastguard Worker       const __m512 vi24x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i24);
674*4bdc9457SAndroid Build Coastguard Worker       const __m512 vk24x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 800);
675*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF, vacc0123456789ABCDEFp0);
676*4bdc9457SAndroid Build Coastguard Worker 
677*4bdc9457SAndroid Build Coastguard Worker 
678*4bdc9457SAndroid Build Coastguard Worker       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
679*4bdc9457SAndroid Build Coastguard Worker       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
680*4bdc9457SAndroid Build Coastguard Worker 
681*4bdc9457SAndroid Build Coastguard Worker       _mm512_mask_storeu_ps(output, vmask, vacc0123456789ABCDEF);
682*4bdc9457SAndroid Build Coastguard Worker       output += c;
683*4bdc9457SAndroid Build Coastguard Worker     }
684*4bdc9457SAndroid Build Coastguard Worker 
685*4bdc9457SAndroid Build Coastguard Worker     output = (float*) ((uintptr_t) output + output_increment);
686*4bdc9457SAndroid Build Coastguard Worker   } while (--output_width != 0);
687*4bdc9457SAndroid Build Coastguard Worker }
688